tradeoffs & constraint · metaphor 14 of 100

You can only feel
the slope.

How does anyone improve at anything, when the only information available is which way is slightly better from here? You cannot see the summit; you can only feel the slope underfoot. That single rule — step downhill — builds skills and also traps them in the nearest valley.

Nobody learning an instrument, a language, or a self can see the global best from where they stand. All you get is local feedback: this practice felt better, that one worse. Follow the improvement gradient faithfully and you will reach a bottom — but the nearest one, which is why diligent people plateau in local optima, why getting better sometimes requires first getting worse.

The whole machinery of learning-by-improvement is one line of arithmetic: measure the slope, take a step against it, repeat. Everything interesting — the plateau, the breakthrough, the person who tries hard and stalls while an erratic beginner sails past — lives in three dials: how big your steps are, where you happened to start, and whether you carry any momentum or noise. Drag the ball. Set it loose.

drag anywhere to set the start · release to let go · height = error, the gap from mastery · green well = the global best
learning rate · step size 0.080

how big a change you make per lesson

noise · stochasticity (SGD) 0.0

varied practice, mistakes, disruption — shakes the ball loose

momentum · inertia

accumulated velocity rolls through small dips toward deeper valleys

scenarios
current error
steps taken
0
slope · |∇|
status
A ball on a hidden skill-landscape. It can feel the slope at exactly one point — its own — and steps against it. Where it ends up is decided before it moves: by the step size, and by where you drop it.
xₜ₊₁ = xₜ − η·∇L(xₜ) + noise v ← μ·v − η·∇L ; x ← x + v Honest gradient descent on a hand-designed error surface L(x). The slope ∇L is the exact analytic derivative; every step, bounce, and escape is computed live, never scripted. η is the learning rate, μ the momentum.

stepping blind

“Just improve from here” always reaches a bottom. Which one is the whole question.

The ball has no map. It cannot see the deep green valley off to the side, cannot compare where it is to where it could be. It knows one number: the slope directly under it — which way, right now, is very slightly better. This is the human condition of learning. You never get to see the space of who you could become. You get feedback on the change you just made, and a direction: this helped, that hurt.

Follow that direction faithfully and you converge. The slope flattens, the improvements shrink, and you come to rest at a place where every small move you can feel makes things worse. That is a real achievement and a real bottom. But local feedback cannot tell a shallow valley from a deep one. At the floor of any basin the slope is zero; the ground goes up in every direction you can test. The plateau of the merely-good and the summit of the truly-great feel identical from the inside. Both say: nothing near here is better. Only one of them is right.

what to try

Sweep the rate. Move the start. Then escape.

  1. Sweep the learning rate through its three regimes. Set it to crawl: the ball inches downhill and, in a whole lifetime of steps, never reaches the floor — agonizing diligence that runs out of time. Set it to just right: smooth, brisk convergence. Set it to too big: the ball overshoots the valley, bounces across it, and flings itself out entirely. Watch the status readout call each one.
  2. Drag the ball to different starting points. Drop it on the left and it falls into a shallow, high valley; drop it a little to the right and the same faithful stepping finds a far deeper one. Nothing changed but the start. Initialization — talent, birth, your first teacher — quietly decides which local optimum all your diligence will find.
  3. Get stuck, then get out. Hit drop me in a trap: the ball settles in a shallow valley and the status reads local min, with a deeper one a ridge away. Raise the noise slider and watch the error tick upward as the ball is shaken out of its comfortable dent — then fall far below where it was stuck. That temporary worsening was the price of the deeper valley. Momentum can't free a resting ball: for that, reset, drop from the far left with momentum on, and watch inertia carry it through the first shallow valley into a deeper one — persistence only helps something already in motion.

local optima are the real subject

Why the disciplined sometimes lose to the erratic.

A diligent person improves faithfully — every step honestly downhill, no wasted motion — and arrives, exactly as promised, at a minimum. They did nothing wrong. Their reward is a plateau: a shallow valley they cannot see out of, because the very rule that got them there (only ever move to something locally better) forbids the temporary climb that a deeper valley would require. Faithful local improvement is precisely the thing that locks in a local optimum. The better you are at never getting worse, the more surely you stay stuck.

Meanwhile the erratic beginner — sloppy, noisy, undisciplined — is being kicked around by their own inconsistency, and every so often that noise flings them over a ridge into a basin the careful person never reached. This is arithmetic: talent is initialization (a lucky starting basin, closer to the deep valley), breakthroughs are noise-driven basin-hops (a disruption that happened to land you somewhere better), and pure diligence optimizes beautifully within whatever basin it began in. The step that feels like backsliding — the botched recital, the year abroad that ruined your technique — is sometimes the only move the geometry allows toward mastery.

the schedule of getting better

Big steps early, small steps late.

You cannot pick one learning rate and keep it. Large steps early are a gift: they let you explore, tolerate bouncing, and cross ridges before you have committed to a valley — the reckless energy of a beginner is functional, not just a phase to be survived. But large steps late are ruin: once you are near a good floor, they only knock you back out. The craft is a schedule — run hot and wide at the start, then cool, shrinking your steps to refine what you have found. A life, run well, does the same: sample widely and cheaply when young, then narrow and deepen.

Momentum is the value of not fully resetting between attempts — of carrying velocity from yesterday's practice into today's, so that a run of consistent effort builds inertia enough to coast through the small dips and plateaus that would stop a fresh start cold. And when the surface is smooth and you are still trapped, the exit is not more careful improvement — it is deliberate disruption: heat the system, accept some worse before you can reach better, then cool again. That schedule of heating and cooling has its own name and page — annealing — and it is the honest answer to when smooth improvement has become a trap.

the mapping

Mathematics ↔ life.

MathematicsLife
the landscape L(x)The space of how-good-you-could-be — every version of the skill or self you might become, laid out as terrain.
height · LError — the gap from mastery. Lower is better; you rarely reach zero.
the gradient · ∇LLocal feedback: which small change helps from exactly where you stand. The only information you ever get.
learning rate · ηHow big a change you make per lesson — crawl, converge, or bounce, decided by this alone.
local minimumThe plateau diligence reaches: a valley where everything nearby is worse, whether or not it is the best there is.
initializationWhere you happened to begin — talent, birth, first teacher — which basin your faithful stepping falls into.
momentum / noisePersistence and disruption: the inertia and the accidents that escape shallow traps pure local stepping can't.

where the metaphor tears

Three honest failures.

The valleys are partly a low-dimensional lie.

This landscape has one dimension, so its valleys are real prisons — walls on both sides. But a skill has thousands of dimensions, and high-dimensional surfaces have astonishingly few true local minima: at almost every flat spot, some direction still leads down. Most "traps" are actually saddle points — passes with a way out you simply haven't found. The dramatic 1-D story of being sealed in a shallow valley is largely an artifact of the picture being drawable. In real learning you are more often lost than imprisoned.

“Get worse to get better” rationalizes thrashing.

The instrument shows a temporary climb in error that pays off in a deeper valley — a strategic basin-hop. But not every regression is one. Most getting-worse is just getting worse: the abandoned method, the destabilizing move that leads nowhere better. The math promises that some uphill steps are necessary; it does not tell you whether yours is a breakthrough in progress or aimless self-sabotage. The comforting story is available to anyone quitting anything.

The surface itself moves — and encodes what "better" means.

Real skill landscapes are non-stationary: the ground shifts as you and the world change, so the valley you were descending may fill in behind you. Worse, the whole shape is set by the loss function — your definition of "better." Optimize a badly-chosen objective perfectly and you arrive, flawlessly, somewhere you never wanted to be; Goodhart lurks in every metric you descend. Choosing the right landscape matters more than descending any landscape well.